Remote MCP server hosting: the options in 2026

TL;DR. A remote MCP server is an MCP server that AI clients reach over HTTPS instead of launching as a local process. ChatGPT, Claude.ai and most cloud-side clients can only use remote servers, so MCP server hosting comes down to four choices: a serverless or edge platform (Cloudflare Workers is the usual example), a general PaaS that runs a long-lived process (Railway, Render), a managed MCP platform that hosts or proxies servers for you, or Hatchable, where every project already exposes its own https://<slug>.hatchable.site/mcp endpoint with OAuth handled and one file per tool, on a free plan with no card.

What is a remote MCP server?

MCP defines two standard transports. With stdio, the client launches the server as a subprocess on the same machine and talks to it over standard input and output. With Streamable HTTP, the server is a web endpoint (a single URL such as https://example.com/mcp) that accepts JSON-RPC over HTTP POST and can stream responses back as server-sent events. Streamable HTTP arrived in the March 2025 revision of the spec and replaced the older HTTP+SSE transport, which is now deprecated. A "remote MCP server" is simply a server that speaks the HTTP transport from a public host.

The distinction matters because of where the client runs. Claude Code, Cursor and Codex run on your machine, so they can launch stdio servers and also call remote ones. ChatGPT and Claude.ai run in their vendors' clouds: as of August 2026, ChatGPT's apps and custom connectors take a URL, not a command, and Anthropic's documentation for custom connectors says the server has to be reachable from the public internet. If you want a tool to work from a browser tab, a phone, or a cloud agent, it has to be remote. If you are new to the protocol, what MCP is covers the basics.

What a host has to provide

Whether you host an MCP server yourself or pay someone to, the same checklist applies.

• HTTPS with a valid certificate. Every cloud client requires it, and the spec tells servers to validate the Origin header and authenticate connections.
• Authentication. Two shapes are common: OAuth (the client discovers your authorization server, the person signs in once in a browser, tokens are issued per client and can be revoked) or a static bearer token in an Authorization header. OAuth is what ChatGPT and Claude.ai expect for anything that touches a real account; bearer tokens are fine for a personal tool pasted into Claude Code or Cursor.
• Long-lived responses. A tool that takes twenty seconds needs the host to keep the HTTP response open, and a server that streams progress needs SSE to pass through proxies intact. Serverless platforms typically cap execution time; check the limit against your slowest tool.
• Logging. Which client called which tool with which arguments, especially once a server is shared.
• State. Most useful tools read or write something. Budget for a database and for secrets storage, because the protocol does not give you either.

The July 2026 revision of the spec, as summarised by Cloudflare in August 2026, moves the protocol further toward stateless plain HTTP and formally deprecates the legacy HTTP+SSE transport, which makes edge and serverless hosting easier than it was a year ago.

MCP server hosting: the four options

1. Serverless and edge platforms

Cloudflare Workers is the best-known example, and Vercel functions work the same way. These platforms are genuinely good at stateless Streamable HTTP: low latency everywhere, near-zero idle cost, and in Cloudflare's case first-party libraries for building the server and wrapping it in OAuth. The trade-off is that you assemble the rest. The OAuth provider still needs an identity source (GitHub, Google, your own), the database is a separate product, and long-running tools run into execution limits. Good fit: a developer who already lives on the platform and wants a public tool for many users.

2. General PaaS (a container or a process)

Railway, Render, Fly.io and similar hosts run any long-lived process, so they can host an MCP server written in any language, including one that wraps a local CLI. Several publish MCP templates; Render's, for example, ships with a generated bearer token, and a Railway template exposes servers over both SSE and Streamable HTTP. TLS, logs and restarts are handled for you. What you own is the app itself: the auth layer, upgrades, the database you attach, and an always-on bill, since these are processes, not functions. Good fit: a Python server, a wrapper around an existing service, or anything that needs to hold a connection open.

3. Managed MCP platforms

A newer category hosts or proxies MCP servers as a service: registries such as Smithery and Glama that run servers for you behind one gateway, and integration platforms such as Composio that expose hundreds of pre-built tools with the OAuth already wired. They are the fastest way to give an agent a known third-party tool (a CRM, an issue tracker) without running anything. The trade-off is that you are usually running their catalogue, on their terms and pricing; custom code is possible on some of them but it lives inside their runtime, and the credentials your agent uses pass through their gateway. Good fit: connecting agents to SaaS you do not control.

4. Hatchable: every project is already a remote MCP server

On Hatchable the MCP endpoint is a property of the project. Create a project, drop one JavaScript file per tool into mcp/, deploy, and it answers at https://<slug>.hatchable.site/mcp with OAuth discovery built in: the person pastes the URL into Claude, Cursor, ChatGPT or Codex, signs in once on hatchable.com, and your tools appear in their tool list. Tokens are scoped to the project and revocable from the console. Each tool's handler gets the same SDK the rest of the app gets, so it can read the project's private Postgres database, send email, fetch a page with the headless browser, or call an API you declared. The free plan includes unlimited private projects and one published app, no card; Builder is $12 a month for unlimited published apps. The trade-off is that tools are JavaScript running in Hatchable's runtime: it is not a Python or Docker host and will not run a long-lived process, so a server that shells out to a local binary belongs on a PaaS. See the MCP feature page and the developer docs.

Decision table

General characterisations based on public information as of August 2026; check each vendor for current details.

The 15-line tool

To make the last option concrete, this is a complete tool on Hatchable. It lets any connected AI save a note into the project's database, and only the project owner can call it.

// mcp/remember.js
export default {
  access: 'admin',
  inputSchema: {
    type: 'object',
    required: ['text'],
    properties: { text: { type: 'string', description: 'what to remember' } },
  },
  handler: async ({ text }, ctx) => {
    const { rows } = await ctx.db.query(
      'insert into thoughts (text) values ($1) returning id', [text]
    )
    return { ok: true, id: rows[0].id }
  },
}

Add a one-line migration for the thoughts table, deploy, and connect it from Claude Code with claude mcp add --transport http notebook https://<slug>.hatchable.site/mcp. The full walkthrough, including the starter you can fork, is at how to build and host an MCP server.

How to choose

Ask three questions. Who is the client? If it is ChatGPT or Claude.ai, the server must be remote and should use OAuth, which rules out a laptop and pushes you toward a platform that handles authorization. What does the tool touch? If it is your own data, pick the option where the database is closest to the tool; if it is someone else's SaaS, a managed platform may already have it. What language is the tool in? JavaScript fits everywhere on this list; Python or a binary points you at a PaaS. If you want hosting, OAuth and the database as one unit, that is the last row of the table.